Auto review requests omitted sessionId, so pi-ai never added the
x-opencode-session header that OpenCode routes require, and every
reviewed tool call failed with 400 MissingSessionID before its body ran.
Review requests now carry the reviewed Session's id and the new
GenerateOptions purpose 'auto-review'. The DeepSeek Session-log upload
skips that purpose, so a review neither uploads the log nor appends an
acceptance record to the Session.
Refs #7150
The shared DeepSeek listing decoder reads an optional per-entry
`badge: { text, tone }`. Text is one literal or translations keyed by
lowercase language id with a required English entry, each trimmed to 1
to 12 characters on one line; tone is neutral, info, success, or
warning, and an absent or unrecognized tone reads as neutral. An absent
or null badge declares none; an unreadable one is dropped with a listing
problem and never hides its model.
resolveListedCatalog keeps the badges of offered models, both DeepSeek
routes attach them through catalogModelInfo, LlmRuntime.listModels
detaches them, and the Host model catalog forwards them as
ModelCatalogModel.badge. resolveModelInfo omits the badge, so no model
request, Session event, or model-facing tool carries it.
Add @deepseek-ai/dsh-agent-model. It provides ctx.agentModel
(resolveForNextStep, select, applyConfiguredRoute) and owns the
modelSelection projection, which the Agent registry no longer registers.
dsh-agent-loop now injects agentModel; the base and SDK-minimal bundles
and the Python SDK runtime include it.
The provider keeps the projection registered until consumers that are
still unloading finish their cleanup. The model-switch notice becomes an
opt-in per-Agent plugin at @deepseek-ai/dsh-agent-model/notice, and
agent/request now also rejects reasoning-effort changes after step
preparation.
Mount the provider in every composition that loads dsh-agent-loop, and
update consumers, the switch-notice snapshot driver, docs, and the
upgrade guide.
Extension failures, unreadable package identities, a blocked session-log
upload, and degraded Messages replay recur on every request. Log the first
occurrence per field, route, module, or blocking event instead.
Co-authored-by: turtle1999 <[email protected]>
Request-extension fields are auxiliary. A contributor whose preparation
throws is now logged and omitted while the other fields are still sent,
and a post-2xx acceptance failure is logged without failing the request.
Package inventory is best effort: a named manifest without a version, or
a bare package whose manifest is missing or unreadable, reports version
"unknown" instead of failing preparation.
`purpose` is a label for replay and transport metadata, so no adapter should
branch on one caller's value. Exact-model metadata now reports
`reasoning.floorEffort`, the least effort a route accepts, and the title
service selects it for its request. `llm-pi-ai` and `llm-deepseek` dispatch
the effort they receive and no longer recognize `session-title`; the
`minimizeReasoning` field is gone. A route with no registered adapter, such
as the recorded replay lanes serve, reports no floor, so the title request
keeps the adapter default instead of failing before dispatch.
Replaces the `minimizeReasoning` field from the previous commit.
`purpose` is a provider-neutral label for replay and transport metadata, so
an adapter must not branch on one caller's value. `GenerateOptions` gains
`minimizeReasoning`, which any auxiliary caller sets when reasoning would
consume the output budget it reserved for a short visible result; both
shipped adapters dispatch such a request at the least reasoning the model
supports, and `session-title-llm` sets it instead of relying on the
adapter recognizing `'session-title'`. The pasted type in the subsystem
page and the generated Cordis API catalog follow.
Prompted by review: llm-pi-ai must not know about session-title.
Session-title calls carry a 64-token output cap and rely on the adapter to
keep reasoning out of it. llm-deepseek maps purpose 'session-title' to
thinking disabled, but llm-pi-ai ignored the purpose, so pi-ai routes sent
title calls at the route's default effort. On HFAI's adaptive-thinking
Claude models (profile reasoning: max) the thinking consumed the cap, the
call finished with max-tokens, and the session kept its fallback title.
The pi-ai adapter now dispatches a session-title request at the model's
lowest supported level: off when the model offers it, otherwise the
lowest declared level. GenerateOptions.purpose documents that obligation
for every adapter.
Optional request extension fields had no size bound. A very large field
could exceed the runtime's string limit while the request was serialized,
which failed every attempt, including retries.
- Bound the extension payload carried by one request (8 MiB by default,
configurable) and send the remainder with later requests.
- If the combined request still cannot be serialized, send the base
request without extension fields and log which fields were omitted.
- Add regression tests and update the related references and decision note.
Log tool changes as developer messages independently of model capability,
fold committed headers and updates into Session.toolHistory(), and let the
LLM runtime construct deferred and retained declarations for routes that
declare a toolUpdate mode. DeepSeek serializes updates as system-role
tool_addition/tool_removal blocks with defer_loading and the tool-changes
beta header. An unchanged restored tool continues the series and is
re-offered through its recorded blocks instead of rewriting declarations.
Replace the obsolete rejection tests with wire assertions and add coverage
for the session fold, loop emission, route projection, config validation,
and the beta header. Chat and Trajectory changes land separately.
Session V4 reserves developer-role messages for incremental model-visible
session changes. The role, tool addition/removal blocks, and deferred-loading
marker need a durable representation before independent provider and UI
integrations can emit or consume them. Known developer events must match an
open turn and step; unsupported providers and UI consumers fail explicitly.
Request headers already capture full tool definitions. Persisting another
ToolSchema in each addition would duplicate Session information and create
two definitions that could drift. Additions therefore store toolName, and
the containing developer event uses headerSeq to select one earlier known
request/header for all its additions. A same-name replacement keeps its
historical definition identity through restart, fork and surface changes.
Removals require no definition copy, and ordinary developer messages omit
headerSeq. Generic sourceEventSeqs remains independent derivation metadata.
The rejected inline tool field is optional never with @persistenceReserved.
The schema extractor retains that uninhabited slot, so permitting a future
value requires a format bump instead of being mistaken for compatible
optional metadata. Unrelated optional fields remain additive and readable.
Native and Session admission reject missing, forward, non-header, unknown,
ambiguous and incomplete referenced definitions, and reject the retired
inline tool field. Validate complete schema fields only for referenced
definitions, preserving unrelated header behavior and extra JSON metadata.
Unknown ignorable records remain opaque. Empty developer nodes retain their
surface position without producing a message or shadowing the system head.
No automatic emitter, provider activation, UI implementation, or speculative
public resolver is added.
Keep the core-owned developer source binding beside the introduced role,
including the request-only type rejection and attribution round trips. Update
consumer guards, SDK/snapshot handling, bilingual owner documentation, and
the same unreleased V4 acknowledgement. Keep earlier tool/source migration
work and its documentation in the parent. Preserve the existing recorded
Cordis inspection generation from the original developer change.
Integration keeps the final source layer as the sole parent and preserves
master's Messages input filtering, empty-user omission, empty tool results,
and system-update placement. Developer/deferred guards still run before
conversion. Native V4 admission, header references, reservation declarations,
and generic tooling match the validated marker candidate. Parent-owned
fixture and generator fixes remain in the parent.
Earlier feature validation passed full build, full lint, 41 doc-sync gates,
and 16 hygiene gates. The unchanged native implementation retains that evidence.
The focused Session/V4/request-input run passes 884 tests across 38 files with
100% statement, branch, function and line coverage for the V3-to-V4 package
and Session surface. Consumer tests pass 482 with one existing skip; tool,
message and optional-field tests pass 146. Cordis inspection and compaction
replays pass in both source and built modes (two cases per mode). Ten invalid
referenced-definition/unknown-header controls failed before their scoped
guards and pass after them.
Reservation validation: the actual generated developer root records tool as
optional primitive never. Widening it requires a bump, while traceId remains
additive and preserved; all 85 focused reader/classifier tests pass.
Final-parent validation: 205 tests across eight provider, type, catalog,
Session and native-format files pass. The preserved parent fixture verifies
Messages input-history replay; runtime/declaration and parent-owned-file
comparisons show no loss during the final rebase.
Co-authored-by: Tianyi Cui <[email protected]>
Give durable user context its producer-owned source kind instead of a generic plugin wrapper. The V3-to-V4 conversion preserves the frozen rename and collision rules, all own JSON metadata, event coordinates, and historical request-only kinds. Native catalog and JSONL admission retain source validation so unknown attribution can survive without its producer while malformed owned sources are refused.
Use an explicit user-only RequestUserInput for auxiliary LLM prompts that do not own durable message identity. Auto Review and compaction remove their request-only source registrations while retaining prompt content and freezing; provider serializers and projections keep the same wire data, assistant replay isolation, image ordering, and tool correlation. Session writes, Agent delivery, and recorded title requests continue to require identified durable messages.
Bind the saved compatibility policy only to user sources and qualify tmux location attribution without weakening fixed system, model, or tool semantics. Rebuild the current inventory and unpublished V4 acknowledgement on the actual tool-role parent, preserving its rationale and all accepted historical records. The catalog format version records tooling metadata; Session remains V4. Developer-role policy and activation belong to their own feature.
The latest-master Messages adapter, serialization, request-input typing, EpochHeader reservation, and attribution checks pass 103 tests across five files. The retained V3 Messages input-history snapshot passes through the source profile; Host/Client typechecks and all 41 documentation gates pass. The generated comparison records four changed roots and 447 unchanged type fingerprints. Latest-master ARIA changes are retained, with producer source labels applied independently; new Messages fixtures retain their producer names as declared source kinds and use the fixed system-message helper; recorded Session generations are unchanged by this regrouping.
Co-authored-by: Tianyi Cui <[email protected]>
Represent tool results as tool-role messages and migrate released wrappers
without losing outer own JSON fields, call identity, error state, or writer
field order. Refuse nested or wrapper-only information that has no lossless
current representation before successor publication.
Validate current event relationships directly instead of projecting V4 data
back into V3. Keep lifecycle, PTC, title/command, compaction, system-field and
retired-syntax checks on catalog restoration and ordinary JSONL opens. The
Worker JSON equality helper must inspect own keys so it cannot admit PTC
arguments that Node rejects.
Keep the source role change, required native admission and consumer fixture
updates in one complete review layer. Explain shared-type changes in the
persistence record and the native reader rationale in its existing note.
Use checked branded fixture identities and discriminant narrowing instead
of unknown double casts; fixture values and assertions remain intact.
This recomposes the original tool-role work by _Kerman with the independently
validated migration and admission repairs by Tianyi Cui. Frozen generations
and snapshot input version selection are retained.
Co-authored-by: Tianyi Cui <[email protected]>
Move the IMAGE_OFFLOAD_REQUIRED recovery out of llm-retry into the new
@deepseek-ai/dsh-compaction-image-offload package, a sibling of the
tool-result pruner: image offload is compaction in another capacity
dimension. Each replaced node is priced through the seam's compaction/prune
shadow-price protocol before its marked copy lands. llm-retry returns to
master; the base bundle mounts the executor.
Drop the image/offload event and every session change with it. An adapter's
IMAGE_OFFLOAD_REQUIRED failure now makes dsh-llm-retry replace the surface
nodes carrying the named oldest images with copies whose blocks are marked
offloaded, through the same surfaceOp: replace compaction uses, then retry the
step. ImageBlock.offloaded becomes a durable field the routes project and the
token meter prices; the session core, its event map, derivation, and
validation return to master.
Remove the llm-image-offload package and every registration it needed. The
recovery branch lives in dsh-llm-retry, which already sits on the
agent/request-error waterfall: an IMAGE_OFFLOAD_REQUIRED failure advances the
image/offload watermark by the count the adapter names and retries the step
outside the provider retry budget. Drop pre-dispatch planning with it: routes
no longer declare a request-image budget on their model info, and the byte
checks that stay in the DeepSeek and pi-ai adapters are provisional until the
provider reports uncacheable images itself.
Keep the agent loop unchanged: the new @deepseek-ai/dsh-llm-image-offload
plugin plans the image/offload watermark on agent/pre-step from the routed
model's budget and advances it on agent/request-error when an adapter fails
with IMAGE_OFFLOAD_REQUIRED, returning the retry action the loop already
honours. Drop PreparedLlmCall.imageRequest and its validation, the planner
wrapper, the replay adapter's imageRequestMaxBytes option, and the duplicate
ACP snapshot scenario; share the exact-byte overflow count between the
DeepSeek and pi-ai adapters through requiredImageOffload(). The SDK image
snapshot replays an authored IMAGE_OFFLOAD_REQUIRED attempt instead of a
replay-only budget.
Use the same logged per-node normalization for an empty rendering as for incapable routes and broken request series: empty each active later system node, then empty the first node. Clearing only the latest node exposes stale earlier instructions, including converter-demoted user text. Dormant empty tails neither override effective text nor produce repeated replacements.
Keep the cleared structural head and ordinary restoration semantics: a capable continuing series may append new nonempty instructions, while incapable or broken-series requests refill the head. No initial empty-head creation change, deletion event, or rollback machinery.
Tests: capable/plain clearing after three active prompt versions fails before the fix and passes afterward; repeated clear and seeded resume remain empty without duplicate system events, restoration includes only new instructions, pi-ai conversion has no stale prompt, and source-event replay reconstructs every request. Final 395 loop tests pass with exact agent.ts/runtime-context.ts coverage at 100% all metrics; source tsc and focused lint pass. Fix new test discriminant narrowing discovered by focused compiler audit. Update EN/ZH consumers and owning rationale, seven pairs verified; parent owns broad gates and generated catalog/SDK/GUI updates.
Yichen Jiang